Back to Research

Open threads

Memory-efficient Adaptation Roadmap

A research thread for comparing LoRA variants under activation memory, quality, and implementation complexity constraints.

Hypothesis

Some memory-saving LoRA variants trade away more update freedom than necessary.

Plan

  1. Build a controlled benchmark.
  2. Track memory, throughput, and validation quality.
  3. Inspect failure modes where the low-rank subspace becomes too restrictive.

Exit criteria

A useful result should explain when each adaptation method is worth using, not only which one wins on a single task.